Intelligent water affair cloud management method and system based on NB-IoT communication technology

Abstract

The application discloses a smart water affair cloud management method and system based on NB-IoT communication technology driving, relates to the technical field of smart water affair monitoring, and comprises the following steps: collecting multi-dimensional real-time sensing data of a pipe network, performing space-time correlation analysis on the pipe network topology model and real-time working condition parameters after standardization processing, and generating a dynamic water affair state correlation graph; inputting the graph into a pre-trained deep learning inference network, performing feature extraction and pattern recognition, outputting the probability distribution of potential abnormal events, and locating the abnormal subgraph structure in the correlation graph, thereby tracing back and determining the key monitoring node or pipe network section causing the abnormality; generating a candidate regulation and control strategy based on historical data, and performing strategy effect deduction and quantitative risk assessment through a simulation environment. The method realizes accurate identification of pipe network abnormalities, spatial positioning of the influence range, and pre-simulation evaluation of regulation and control strategies by constructing a correlation graph model and applying graph deep learning.

Description

Technical Field

[0001] This invention belongs to the field of smart water monitoring technology, specifically a smart water cloud management method and system driven by NB-IoT communication technology. Background Technology

[0002] In the field of smart water management, existing technologies generally rely on NB-IoT sensors deployed at various nodes of the pipeline network to collect data and set independent thresholds for over-limit alarms, or perform simple multi-parameter linkage judgments based on limited rules. These methods treat monitoring points as independent information sources, mainly focusing on the instantaneous state or simple statistical characteristics of their own data. Conventional anomaly detection algorithms are also mostly for analyzing time-series data from a single sensor, or for processing multiple independent data streams in parallel before performing logical combination judgments.

[0003] The limitation of existing technologies lies in their inability to effectively characterize the dynamic transmission and interrelationships of state parameters such as pressure, flow rate, and water quality within the spatial topology of a complex physical system like a water supply network. When a leak, pipe burst, or abnormal water usage event occurs at a point in the network, the resulting state disturbances propagate along specific paths within the network topology, affecting multiple related monitoring points. Existing discrete monitoring methods cannot construct such dynamic spatial impact chains, resulting in the system only being able to detect local data anomalies but struggling to accurately determine the nature and scope of the abnormal event, let alone quickly trace back to the root cause node or section that triggered the chain reaction. Faced with massive amounts of sensor data, maintenance personnel need to rely heavily on extensive manual correlation analysis based on experience, which is inefficient and prone to misjudgment.

[0004] A technical method is needed to organically integrate discrete multidimensional sensor data based on the physical connection of the pipeline network and the real-time hydraulic status to form a systematic dynamic correlation model, and to realize the automatic identification and accurate spatial positioning of abnormal patterns based on this model. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a smart water management cloud method based on NB-IoT communication technology, comprising: Collect multidimensional real-time sensor data on the operation status of the pipeline network, and obtain a standardized water status dataset based on the multidimensional real-time sensor data; Based on the preset pipeline network topology model and real-time operating parameters, a spatiotemporal correlation analysis is performed on the standardized water status dataset to generate a water status correlation diagram that reflects the dynamic correlation within the pipeline network. The water status association map is input into a pre-trained deep learning inference network. The deep learning inference network performs feature extraction and pattern recognition on the water status association map and outputs the inference results of potential abnormal events in the pipeline network and their probability distribution. Based on the inference results, the abnormal subgraph structure associated with potential abnormal events is located in the water status association diagram, and the root cause analysis is performed on the abnormal subgraph structure to determine the key monitoring node or pipeline section that caused the abnormality. Based on historical data from the key monitoring nodes or pipeline sections, several candidate control strategies are generated for the potential abnormal events. The execution effect of the candidate control strategies is simulated using a simulation environment. Based on the simulation results, the effectiveness of each candidate control strategy is evaluated and the risk is quantified, generating a strategy evaluation report.

[0006] Furthermore, the obtained standardized water status dataset includes: The multidimensional real-time sensing data is uploaded to the cloud platform via the NB-IoT network; The cloud platform receives the multidimensional real-time sensing data and performs format parsing and outlier cleaning on the data to obtain a standardized water status dataset, specifically including: The heterogeneous data formats generated by different models of NB-IoT terminal devices in the multidimensional real-time sensing data are identified, and the heterogeneous data formats are converted into a unified standardized data format within the cloud platform according to a preset format mapping table. After the sensor data stream is converted to a standardized data format, a time window-based integrity check is performed, and missing data points are marked and supplemented. Perform outlier detection based on statistical distribution and physical constraints on the sensor data stream with supplemented data points to identify outlier data points that deviate from the normal range or change pattern; For the identified abnormal data points, an interpolation algorithm based on the characteristics of its neighboring data points and historical data from the same period is used to repair the data or remove the flags. The sensor data stream, after data repair or label removal, is subjected to dimensional normalization to convert data values ​​of different physical dimensions into a unified numerical range, and finally outputs the standardized water status dataset.

[0007] Furthermore, based on a preset pipeline topology model and real-time operating parameters, the standardized water status dataset is subjected to spatiotemporal correlation analysis to generate a water status correlation diagram reflecting the dynamic correlation within the pipeline network, including: Load the pipeline topology model that describes the location of monitoring nodes, pipeline connections, and hydraulic properties in the pipeline network; Extract the status data of each monitoring node at the same time from the standardized water status dataset, and assign the status data as an attribute to the corresponding node in the pipeline topology model; Based on the real-time operating parameters, the weights of the edges connecting each node in the pipeline topology model are dynamically calculated, and the weights reflect the real-time strength of the hydraulic connection between the nodes. Construct a graph data structure with monitoring nodes as vertices and weighted edges representing the dynamic relationships between nodes as the initial relationship graph; Within a preset time window, the propagation path and delay effect of node state changes in the initial association graph are analyzed, the weights of the corresponding edges are strengthened or weakened, and finally the water affairs state association graph that can reflect the state propagation and dynamic coupling relationship is generated.

[0008] Furthermore, the water status association map is input into a pre-trained deep learning inference network. The deep learning inference network performs feature extraction and pattern recognition on the water status association map, and outputs the inference results of potential abnormal events in the pipeline network and their probability distribution, including: The attribute features of the nodes and the weight features of the edges in the water status association graph are jointly encoded into a high-dimensional feature tensor. The feature tensor is input into the graph convolutional layer of the deep learning inference network. The graph convolutional layer learns and extracts the local topological features of the water status association graph by aggregating the features of the node itself and its neighboring nodes. The extracted local topological features are input into the attention mechanism layer of the deep learning inference network. The attention mechanism layer calculates the importance weights of different nodes and edges in the graph to the overall state representation, and performs weighted fusion of features accordingly. The weighted and fused features are input into the classification inference layer of the deep learning inference network. The classification inference layer outputs the probability distribution of the water status association graph corresponding to various predefined abnormal event categories. The probability distribution is the inference result.

[0009] Furthermore, based on the inference results, the abnormal subgraph structure associated with potential abnormal events is located in the water status correlation diagram, and root cause analysis is performed on the abnormal subgraph structure to determine the key monitoring nodes or pipeline sections that cause the anomalies, including: Extract potential abnormal event types with a probability exceeding a preset threshold from the inference results; In the water status association diagram, based on the feature patterns corresponding to the potential abnormal event types, a subgraph structure matching the feature patterns is searched, and the searched subgraph structure is marked as a candidate abnormal subgraph. For each candidate anomaly subgraph, the anomaly contribution of its internal node states is calculated, whereby the anomaly contribution quantifies the magnitude of the contribution of node state changes to the overall anomaly pattern. The nodes in the candidate anomaly subgraph are sorted according to the anomaly contribution degree, and the nodes with high contribution degrees are traced upstream or to the root source along the connection edges of the water status association graph. By combining the hydraulic transmission direction of the pipeline network with historical anomaly records, nodes that can serve as the origin of events or continuous pipe segments consisting of multiple nodes are identified in the candidate anomaly subgraph, and these are determined as the key monitoring nodes or pipeline segments.

[0010] Furthermore, by combining historical data from the key monitoring nodes or pipeline sections, several candidate control strategies are generated for the potential abnormal events, including: Retrieve the status data sequence of the key monitoring nodes or pipeline sections under similar past operating conditions and the records of the control actions taken from the historical database; Using the aforementioned pipeline physical model, the transmission effect on the hydraulic state of the entire pipeline network when different control actions are applied at the key monitoring nodes or pipeline sections is simulated. Based on the historical data sequence, the records of control actions, and the simulation results of the pipeline physical model, a strategy knowledge base is constructed. Based on the characteristics of the potential abnormal events, retrieve the basic strategy template from the strategy knowledge base; The retrieved basic strategy template is parametrically mutated and logically combined to generate multiple candidate control strategies that differ in control objectives, control intensity, and execution timing.

[0011] Furthermore, the simulation environment is used to extrapolate the execution effects of the candidate control strategies, and based on the extrapolation results, the effectiveness of each candidate control strategy is evaluated and the risk is quantified, generating a strategy evaluation report, including: Construct a high-fidelity simulation environment based on the aforementioned pipeline network physical model and the current real-time water status data; Each of the candidate control strategies is converted into a sequence of instructions recognizable by the simulation environment and executed sequentially in the simulation environment; During execution, the trajectory of changes in key state variables in the simulation environment is recorded until the preset future time point is deduced. Based on the state of the pipeline network at the end of the simulation, multiple performance indicators for each candidate control strategy are calculated, including the degree of anomaly mitigation, energy consumption change, and recovery time. Identify secondary risk events that occur during the simulation process, and quantitatively assess the probability of occurrence and severity of impact for each identified secondary risk event; The quantitative evaluation results of all the effectiveness indicators and secondary risk events of each candidate control strategy are summarized and compiled into a structured strategy evaluation report.

[0012] Furthermore, the method includes: Based on the strategy evaluation report, an optimal control strategy is selected from the candidate control strategies using a multi-objective decision-making algorithm. The optimal control strategy is then decomposed into a specific set of control commands. The set of control commands is then sent to the field actuators corresponding to the key monitoring nodes or pipeline sections, driving the field actuators to execute the set of control commands to adjust the pipeline operation status. After the field actuator performs the control, it collects new multi-dimensional real-time sensing data through the corresponding NB-IoT terminal device and feeds the new multi-dimensional real-time sensing data back to the cloud platform to start a new round of management cycle; Based on the strategy evaluation report, an optimal control strategy is selected from the candidate control strategies using a multi-objective decision-making algorithm, including: Extract multiple performance indicators and risk quantification values ​​for each candidate control strategy from the strategy evaluation report; The performance index values ​​and risk quantification values ​​are normalized to eliminate the influence of different dimensions and uniformly converted into benefit-type or cost-type indicators. Based on the preference weights preset by the management side, a comprehensive evaluation function is constructed, which is the weighted sum of the normalized index values; Substitute the index value of each candidate regulation strategy into the comprehensive evaluation function to calculate its comprehensive score; Introduce constraints to exclude candidate control strategies that violate key physical limits or safety red lines; Among the candidate control strategies that meet the constraints, the strategy with the highest comprehensive score is selected as the optimal control strategy.

[0013] Furthermore, the optimal control strategy is decomposed into a specific set of control commands, and the set of control commands is issued to the field actuators corresponding to the key monitoring nodes or pipeline sections, including: The optimal control strategy is analyzed and broken down into a series of atomic control actions arranged in a time or logical order; For each atomic control action, match the field actuator of its target and determine the unique network identifier and control interface protocol of the field actuator; The parameters of each atomic control action are encapsulated into a specific instruction data packet according to the control interface protocol of the corresponding field actuator; All instruction data packets corresponding to atomic control actions are arranged into an ordered instruction list according to the execution sequence to form the control instruction set; Through the cloud platform and NB-IoT network, the instruction data packets in the control instruction set are sequentially sent to the corresponding field actuators.

[0014] Furthermore, the present invention also includes a smart water management cloud system driven by NB-IoT communication technology. The system includes a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it implements the smart water management cloud system driven by NB-IoT communication technology as described above.

[0015] Compared with the prior art, the beneficial effects of the present invention are: By combining collected multidimensional real-time sensor data with a pre-defined pipeline network topology model and real-time operating parameters, spatiotemporal correlation analysis is performed to generate a dynamic water status correlation diagram. This diagram maps the physical pipeline network into a data structure containing node states and edge association weights. It not only includes real-time readings of each monitoring point but also encodes the possible propagation paths and mutual influence strengths of state parameters within the pipeline network. This integrates previously isolated sensor readings into a unified network view that reflects the true physical constraints of the system, upgrading the data analysis unit from "points" to "substructures composed of points and their relationships." This technical solution enables the system to perceive the dynamic linkage and propagation effects of the internal state of the pipeline network in a panoramic way, providing an accurate data foundation for capturing complex anomaly patterns caused by single events and spreading across the topological network.

[0016] A pre-trained deep learning inference network is used to directly extract features and recognize patterns in water status correlation graphs. This network can learn high-order nonlinear patterns formed by node features and edge relationships in the graph, identifying subgraph structures that deviate from normal correlation patterns. Its output not only includes anomaly probabilities but also directly locates anomalous subgraph structures strongly correlated with anomalous events. This technical solution enables the analysis process to automatically focus on local areas of abnormal behavior in the correlation network, directly and automatically linking the anomaly inference results with specific pipeline spatial locations. This achieves a shift from "alarming an anomaly at a certain data point" to "locating an anomalous subnetwork and inferring its core location," reducing the time and manual analysis costs required from anomaly detection to locating suspected root causes, and providing a clear automated target for root cause tracing analysis. Attached Figure Description

[0017] Figure 1This is a flowchart illustrating the steps of the smart water management cloud method based on NB-IoT communication technology described in this invention. Figure 2 A flowchart generated for the water status association diagram. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1The overall implementation scheme of the smart water management cloud method driven by NB-IoT communication technology is as follows: Sensing devices integrated with NB-IoT communication modules, deployed at key nodes of the water supply network, periodically or triggered to collect multi-dimensional real-time sensor data, including pressure, flow rate, and water quality parameters. This sensor data is transmitted to the cloud management platform via a low-power wide-area NB-IoT network. The cloud management platform receives multi-dimensional real-time sensor data from a massive number of terminals, cleans, parses, and normalizes it to form a standardized water status dataset. The platform loads a pre-built network topology model containing the spatial location of network nodes, pipeline connection relationships, and physical attributes, and combines this with real-time operating parameters such as pump start / stop and valve opening to perform spatiotemporal correlation analysis on the standardized water status dataset, constructing a water status correlation graph with monitoring nodes as vertices and dynamically weighted edges reflecting the hydraulic correlation strength between nodes. This correlation graph is input into a deep learning inference network pre-trained on a large amount of historical data. This network extracts features and recognizes patterns from the input graph through its graph convolutional layers and attention mechanisms, outputting probability distribution inferences of the current state of the pipeline network corresponding to various potential abnormal events (such as leaks, pipe bursts, and water pollution). Based on the high-probability abnormal event inference, the system performs subgraph matching and searching within the water status correlation graph to locate abnormal subgraph structures matching the abnormal patterns. By calculating the abnormal contribution of nodes and tracing back along the pipeline network topology, it identifies one or more key monitoring nodes or pipeline sections that triggered the abnormal events. The system then retrieves historical data and control records of the key node or section under similar operating conditions from the historical database. Combining this with a pipeline network hydraulic model, it simulates the transmission effects of different control actions, generating multiple candidate control strategies that differ in control objectives, intensity, and timing. Each candidate control strategy is imported into a simulation environment based on a high-fidelity pipeline network physical model and real-time state initialization for performance evaluation. The simulation process records changes in pipeline network state variables and calculates various performance indicators after strategy implementation. It also identifies and quantifies potential secondary risks that may arise during the simulation. Finally, the system generates a strategy evaluation report containing performance indicators and risk quantification results for each candidate strategy, providing support for management decision-making.

[0020] In one embodiment of the present invention, in a specific implementation, the cloud platform receives multi-dimensional real-time sensing data uploaded through the NB-IoT network. This multi-dimensional real-time sensing data originates from various models of NB-IoT terminal devices deployed in the network. The cloud platform initiates a data standardization processing flow. The flow identifies heterogeneous data formats generated by different models of NB-IoT terminal devices in the multi-dimensional real-time sensing data. The heterogeneous data formats are reflected in the differences in data packet structure, byte order, and encoding method. The cloud platform converts the heterogeneous data formats into a unified standardized data format within the cloud platform according to a format mapping table pre-configured in the database. The format mapping table defines the mapping relationship between each device model identifier and standard field names, data types, and storage precision.

[0021] In practical implementation, a time-window-based integrity check is performed on the sensor data stream after it has been converted to a standardized data format. The time window is set to a fixed duration based on the data reporting cycle. The integrity check is performed by comparing the number of data points expected to be received within the time window with the number of data points actually received. For marked missing data points, a time-series linear interpolation method is used to fill in the missing data points using adjacent valid data points. In some embodiments, the length of the time window can be dynamically adjusted according to the real-time requirements of pipeline monitoring. The time window is shortened during high-frequency monitoring and extended during low-frequency monitoring. The integrity check also checks the continuity of the timestamps of the data points and marks data points with timestamp jumps or duplicates. In some embodiments, for continuously missing data points due to communication interruptions, the filling algorithm switches to using the average of the data series from the same historical period to fill the missing data. The historical data is taken from the aggregated statistical data of the same days of the week and the same time in the past from the historical database.

[0022] Optionally, the outlier detection process simultaneously applies detection methods based on statistical distribution and physical constraints. Statistical distribution-based detection calculates the upper quartile, lower quartile, and interquartile range for each monitoring parameter's data stream. Data points with values ​​higher than the upper quartile plus 1.5 times the interquartile range or lower than the lower quartile minus 1.5 times the interquartile range are initially identified as statistical outliers. Optionally, physical constraint-based detection sets hard boundaries for each monitoring parameter based on pipeline design parameters and fluid mechanics principles. For example, the effective range for pressure monitoring values ​​is set between the minimum service pressure and the maximum pressure-bearing capacity of the pipeline network, and the effective range for flow monitoring values ​​is set between zero and the maximum transport capacity of the pipeline. Any data point exceeding these hard boundaries is identified as a physical outlier.

[0023] It is understandable that identified abnormal data points enter the data repair or removal stage. The repair process employs an interpolation algorithm based on the characteristics of its neighboring data points and historical data from the same period. The algorithm prioritizes using linear interpolation with two valid data points before and after the abnormal data point to calculate a replacement value. It is also understandable that when an abnormal data point is at the end of a sequence and lacks adjacent valid data points, the interpolation algorithm references the data distribution characteristics of the same historical time period for that monitoring point, using a weighted average of historical data for repair. If none of the repair conditions are met, the data point is marked for removal. In specific implementation, the sensor data stream after data repair or removal is subjected to dimensional normalization. Dimensional normalization converts data values ​​with different physical units, such as pressure, flow rate, and turbidity, into a unified numerical range. The normalization process uses a minimum-maximum scaling method. For any monitoring parameter j, its normalized value... The result is obtained through calculation using the formula: in: This represents the original value of parameter j in the i-th data point. This indicates the minimum value of parameter j among all data points in the current processing batch. This indicates the maximum value of parameter j among all data points in the current batch. These are normalized values ​​falling within the [0,1] interval. All parameters, after normalization, collectively constitute a standardized water status dataset. In practice, the query and conversion operations of the format mapping table are executed in real-time by the rule engine in the cloud platform. The rule engine matches the device identifier code in the data packet and calls the corresponding conversion function. The conversion function parses the original data fields and reassembles them into standardized data objects with a unified structure. Throughout the implementation, the entire data processing flow generates detailed log records, including data reception time, format conversion results, details of missing data supplementation, outlier detection reports, and normalization parameters. These log records are stored in the cloud platform database for auditing and backtracking purposes.

[0024] In one embodiment of the present invention, see [reference] Figure 2In the specific implementation, a pipeline topology model describing the location of monitoring nodes, pipeline connection relationships, and hydraulic properties in the pipeline network is loaded. The pipeline topology model is stored in the cloud platform database in the form of a graph data structure. In the graph, vertices correspond to monitoring nodes in the physical pipeline network, and edges correspond to pipeline segments connecting monitoring nodes. The status data of all monitoring nodes at the same acquisition time are extracted from the standardized water status dataset. The status data includes pressure values, flow values, and water quality parameter values. These status data are assigned as attribute features to the corresponding vertices in the pipeline topology model. The weights of the edges connecting each node in the pipeline topology model are dynamically calculated based on real-time operating parameters, including pump start / stop status and valve opening ratio. The weights reflect the real-time strength of the hydraulic connection between nodes. A graph data structure with monitoring nodes as vertices and weighted edges representing the dynamic association between nodes is constructed as the initial association graph.

[0025] In practice, the propagation path and delay effect of node state changes in the initial correlation graph are analyzed within a preset time window. The time window is set to five minutes. The analysis process continuously monitors the pressure value sequence of each node. When a significant change in the pressure of a node is detected, the path and time delay of the change propagating along the weighted edges to adjacent nodes in the initial correlation graph are tracked. If the pressure change of a node causes a regular change in the pressure of adjacent nodes within the expected time, the weight of the edge connecting the two nodes is increased. The increase is proportional to the correlation coefficient of the pressure change. If the expected propagation effect is not observed, the weight of the corresponding edge is weakened. The degree of weakening is based on the statistics of historical propagation success rate. Through this dynamic adjustment, a water status correlation graph that reflects the relationship between state propagation and dynamic coupling is finally generated.

[0026] In practice, the water status association graph is input into a pre-trained deep learning inference network. The deep learning inference network encodes the attribute features of nodes and the weight features of edges in the water status association graph into a high-dimensional feature tensor. This feature tensor is then input into the graph convolutional layer of the deep learning inference network. The graph convolutional layer learns and extracts local topological features of the water status association graph by aggregating the features of the node itself and its neighbors. The graph convolution operation is performed on each vertex of the graph. For vertex v, its new feature representation... The result is calculated using the formula: in: This represents the feature vector of vertex v at the l-th layer. This represents the set of neighboring vertices of vertex v in the water status association graph. It is a normalization constant. It is the trainable weight matrix of the l-th layer. Representing a nonlinear activation function, through the stacking of multiple graph convolutional layers, the deep learning inference network can capture information about multi-hop neighbors in the water status association graph. In specific implementation, the extracted local topological features are input into the attention mechanism layer of the deep learning inference network. The attention mechanism layer calculates the importance weights of different nodes and edges in the graph to the overall state representation, and performs weighted fusion of features accordingly. For any two vertices p and q in the graph, the attention mechanism layer calculates an attention coefficient, which indicates the importance of the features of vertex q to vertex p. The weighted fused features of vertex p are the weighted sum of the attention coefficients of the features of all its neighboring vertices. In specific implementation, the weighted fused features are input into the classification inference layer of the deep learning inference network. The classification inference layer consists of a fully connected neural network layer and a softmax output layer. The softmax output layer converts the output of the fully connected layer into a probability distribution, which represents the probability of the water status association graph corresponding to various predefined abnormal event categories. The output probability distribution is the inference result of the potential abnormal events of the pipeline network and their probability distributions.

[0027] In one embodiment of the present invention, based on the inference results of potential abnormal events and their probability distributions in the pipeline network output by the deep learning inference network, the system extracts potential abnormal event types with probability values ​​exceeding a preset threshold from the inference results of the probability distribution. The preset threshold is set to 70%, and event types with probability values ​​exceeding 70% are identified as high-probability potential abnormal event types. In a specific implementation, the system searches for subgraph structures that match the feature patterns corresponding to the potential abnormal event types in the water status association graph. The feature patterns corresponding to the potential abnormal event types are learned by the deep learning inference network during the training phase and stored in the form of template subgraphs. The search and matching process performs subgraph isomorphic matching between the water status association graph and the template subgraphs, calculates the similarity scores between each connected subgraph in the water status association graph and the template subgraph in terms of node attribute distribution and edge weight patterns, and marks connected subgraph structures with similarity scores higher than the matching threshold as candidate abnormal subgraphs.

[0028] In some embodiments, template subgraphs are stored as a form containing typical anomalous topology and state feature vectors. For example, the characteristic pattern of a pipeline leak event might manifest as a star-shaped structure with a central node experiencing a sudden pressure drop and its adjacent nodes experiencing a gradual pressure drop. The search algorithm searches for subgraphs in the graph that conform to this structure. In some embodiments, the matching process allows for a certain degree of flexible matching, meaning that a match can still be considered successful if node attribute values ​​fluctuate within an allowable error range or if the subgraph size differs slightly. In a specific implementation, the anomalous contribution of the internal node states is calculated for each candidate anomalous subgraph. The anomalous contribution measures the magnitude of the contribution of node state changes to the overall anomalous pattern. The anomalous contribution of node k is... The result is calculated using the formula: in: This represents the Euclidean distance between the current state attribute vector of node k and the historical normal baseline state vector. The larger the distance, the further it deviates from the normal state. This represents the betweenness centrality of node k in the candidate anomaly subgraph, measuring the importance of node k in the state propagation path within the subgraph. This is a preset balancing weight coefficient used to adjust the relative importance of state deviation and topological centrality in contribution calculation. Optionally, the historical normal baseline state vector is obtained by statistically analyzing the mean state data of the node during historical anomaly-free periods. In specific implementation, nodes in the candidate anomaly subgraph are sorted in descending order according to their anomaly contribution, resulting in a node sequence. In specific implementation, the process traces upstream or towards the root source from the node with the highest anomaly contribution along the connecting edges in the water status association graph. The tracing process starts from the node with the highest anomaly contribution, queries the incoming edges in the water status association graph pointing to that node (i.e., the upstream connections in the water flow direction), jumps to the upstream node, and compares the anomaly contribution and state change timestamp of the upstream node with those of the current node. If the upstream node also has a high anomaly contribution and its state change time is earlier than that of the current node, then the upstream node is included in the root source tracing path. It can be understood that tracing is an iterative process, repeating the above jump and comparison operations until a node is found whose anomaly contribution of all its upstream nodes is significantly lower than that node's or whose state change time is significantly later than that node's, at which point the tracing process terminates. Optionally, the selection of the tracing direction is determined by combining the water flow direction calculated by the real-time hydraulic model to ensure that the tracing is carried out in the reverse direction of the actual water flow.

[0029] It is understandable that by combining the hydraulic transmission direction of the pipeline network with historical anomaly records, nodes that can serve as the origin of events or continuous pipe segments consisting of multiple nodes in the candidate anomaly subgraph can be identified. Hydraulic transmission direction information is obtained from the pipeline network topology model, and historical anomaly records are queried from the historical event database. If a node on the current tracing path frequently serves as the origin point of similar events in historical anomaly records, the confidence level of its identification as a key monitoring node is increased. In specific implementation, the tracing termination point—that is, the single monitoring node most likely to be the origin of the event, or the pipeline segment consisting of multiple continuous nodes and their connecting edges on the tracing path—is identified as the key monitoring node or pipeline segment that triggered this potential anomaly. The information of the key monitoring node or pipeline segment, along with its anomaly contribution ranking, is output together with the tracing path records.

[0030] In one embodiment of the present invention, in a specific implementation, several candidate control strategies for potential abnormal events are generated by combining historical data of key monitoring nodes or pipeline sections. This process retrieves the state data sequence of key monitoring nodes or pipeline sections under similar past operating conditions and the records of control actions taken from the historical database. The historical database stores the time-series data of all monitoring nodes and all control operation logs during the long-term operation of the pipeline network. The judgment of similar operating conditions is based on the shape of the daily water consumption curve, seasonal characteristics, and the matching degree of day and night time periods. In a specific implementation, the pipeline physical model is used to simulate the transmission effect on the hydraulic state of the entire pipeline network when different control actions are applied to key monitoring nodes or pipeline sections. The pipeline physical model is a high-fidelity mathematical model constructed based on the pipeline topology, pipeline hydraulic properties, and fluid dynamics equations. In the simulation process, control actions such as valve opening setpoints or pump station start / stop commands are input as boundary conditions into the pipeline physical model. By solving the equations, the spatiotemporal distribution of pressure and pipeline flow of all nodes in the pipeline network during the simulation period is obtained.

[0031] A strategy knowledge base is constructed based on historical data sequences, control action records, and simulation results of the pipeline physical model. The strategy knowledge base stores entries in the form of database tables. Each entry is associated with a historical abnormal scenario, the combination of control actions taken, the transmission effect prediction results of the pipeline physical model simulation, and the post-control state response observed in the actual system. See Table 1 for an example of entries in the strategy knowledge base.

[0032] Table 1: List of Items in the Strategy Knowledge Base In practice, basic policy templates are retrieved from the policy knowledge base based on the characteristics of the current potential abnormal events. The retrieval process encodes the characteristics of the current potential abnormal events into feature vectors. The feature vectors include the abnormality type code, the key monitoring node location code, the abnormality severity level, and the occurrence timestamp. The cosine similarity between the feature vector and the feature vector associated with the abnormal scene identifier of each entry in the policy knowledge base is calculated. Basic policy templates with similarity exceeding a preset threshold are selected as the retrieval results.

[0033] In some embodiments, the preset threshold is set to 0.8, and the similarity calculation ignores the absolute date in the timestamp but considers periodicity and time period information. The retrieved basic strategy templates are parametrically mutated and logically combined to generate multiple candidate control strategies that differ in control targets, control strengths, and execution sequences. Parametric mutation refers to applying random offsets or adjusting the control action parameters of the basic strategy template by a fixed step size, such as mutating the valve opening adjustment value from 50% to 40% or 60% of the basic template. Logical combination refers to sequentially connecting or simultaneously parallelizing the execution steps of multiple basic strategy templates to form a new composite strategy sequence. In some embodiments, parametric mutation uses a Gaussian perturbation method to sample continuous parameters such as valve opening near their original values ​​according to a normal distribution to generate new values. Optionally, logical combination is represented by a directed acyclic graph, where nodes represent atomic control actions, and edges represent sequential dependencies or concurrent relationships between actions.

[0034] In practical implementation, a simulation environment is used to extrapolate the execution effects of candidate control strategies. Based on the extrapolation results, the effectiveness of each candidate control strategy is evaluated and the risks are quantified to generate a strategy evaluation report. A high-fidelity simulation environment is constructed based on the pipeline network physical model and the current real-time water status data. The high-fidelity simulation environment uses the latest status data of all monitoring nodes extracted from the standardized water status dataset as the initial condition to load the same pipeline network physical model as in the strategy extrapolation stage. Each candidate control strategy is transformed into a sequence of instructions recognizable by the simulation environment and executed sequentially within the simulation environment. The instruction sequence is a set of control commands with absolute or relative timestamps. The simulation environment engine applies the control commands to the corresponding actuator nodes of the pipeline network physical model at the corresponding time points.

[0035] During execution, the changes in key state variables in the simulation environment are recorded until a preset future time point is reached. Key state variables include pressure, flow, and water quality parameters at all monitoring nodes. The change trajectories are stored in memory as time-series arrays for subsequent analysis. The preset future time point is dynamically set based on the expected handling time for potential abnormal event types; for sudden pipe bursts, the preset future time point is two hours later, and for slow water quality deterioration events, it is twenty-four hours later. Based on the network status at the end of the simulation, multiple effectiveness indicators for each candidate control strategy are calculated. These indicators include the degree of anomaly mitigation, energy consumption change, and recovery time. The result is calculated using the formula: in: This represents the set of monitoring nodes affected by the anomaly. This represents the pressure value of node i at the start of the simulation. This represents the pressure value of node i at the end of the simulation. This represents the baseline pressure value of node i under normal conditions. The energy consumption change is obtained by calculating the difference between the total power consumption of all pump stations during the simulation period and the power consumption of the historical baseline operating conditions. The recovery time is defined as the longest time from the start of the simulation until the pressure values ​​of all affected monitoring nodes continuously recover to the allowable error range of their normal baseline pressure values.

[0036] In practical implementation, secondary risk events that occur during the simulation process are identified, and the probability of occurrence and severity of impact of each identified secondary risk event are quantitatively assessed. Secondary risk events include pressure exceedance events, flow velocity exceedance events, and water quality index deterioration events. The probability of occurrence is calculated based on the proportion of the event's duration within the simulation time window to the total duration. The severity of impact is calculated by comprehensively considering the geographical scope of the event's impact, the number of affected nodes, and the extent to which the status deviates from normal values, resulting in a severity score. For example, a pressure exceedance event refers to a node pressure exceeding the safe upper limit or falling below the minimum service pressure, and a flow velocity exceedance event refers to a pipeline flow velocity exceeding the maximum allowable design value. All effectiveness indicators and the quantitative assessment results of secondary risk events for each candidate control strategy are compiled into a structured strategy assessment report. The strategy assessment report lists each candidate control strategy's number, anomaly mitigation level, energy consumption change, recovery time, and a list of secondary risk events in tabular form. Each event in the secondary risk event list includes the event type, probability of occurrence, and severity score.

[0037] In one embodiment of the present invention, in a specific implementation, an optimal control strategy is selected from candidate control strategies based on a strategy evaluation report using a multi-objective decision-making algorithm. The selection process extracts multiple performance index values ​​and risk quantification values ​​for each candidate control strategy from the strategy evaluation report. Performance index values ​​include the degree of anomaly mitigation, energy consumption change, and recovery time. Risk quantification values ​​include the probability of secondary risk events and the severity of their impact. The performance index values ​​and risk quantification values ​​are normalized to eliminate the influence of different dimensions and uniformly convert them into benefit-type or cost-type indicators. The normalization process uses a minimum-maximum scaling method to map the original values ​​of all indicators to the [0,1] interval. A comprehensive evaluation function is constructed based on the preset preference weights of the management side. The comprehensive evaluation function is a weighted sum of the normalized indicator values ​​and is used to calculate the comprehensive score of each candidate control strategy. in: This represents the overall score. This represents the normalized index value indicating the degree of abnormal mitigation. This represents the normalized energy consumption change index value. This represents the normalized recovery time index value. This represents the normalized comprehensive risk index value, which is calculated by combining the probability of occurrence and severity scores of each secondary risk event. It is the preference weight preset by the management side and satisfies The normalized index value of each candidate control strategy is substituted into the comprehensive evaluation function to calculate its comprehensive score. Constraints are introduced to exclude candidate control strategies that violate key physical limits or safety red lines. The constraints include that the pressure at any node during the simulation must not be lower than the minimum service pressure and the flow velocity in any pipeline must not exceed the maximum design flow velocity. Candidate control strategies that violate any constraint are directly excluded. Among the candidate control strategies that meet the constraints, the strategy with the highest comprehensive score is selected as the optimal control strategy.

[0038] In practical implementation, the optimal control strategy is decomposed into a set of specific control commands, which are then distributed to the field actuators corresponding to key monitoring nodes or pipeline sections. The decomposition process analyzes the optimal control strategy, breaking it down into a series of atomic control actions arranged in a time- or logical sequence. These atomic control actions include basic commands such as "adjusting the valve to a specific opening," "starting or stopping the water pump," and "opening or closing the pressure relief valve." Each atomic control action is matched with its target field actuator, and a unique network identifier and control interface protocol are determined for each actuator. The unique network identifier of the field actuator is the device code of its integrated NB-IoT communication module, and the control interface protocols include standard industrial protocols such as Modbus and MQTT. The parameters of each atomic control action are encapsulated into specific instruction data packets according to the control interface protocol of its corresponding field actuator. The encapsulation process assembles data frames containing the target address, function code, parameter register address, parameter value, and cyclic redundancy check code according to the protocol specifications. All instruction data packets corresponding to the atomic control actions are arranged into an ordered instruction list based on their execution sequence, forming a control command set. This instruction list is a queue with a global timestamp or relative delay time.

[0039] The cloud platform and NB-IoT network sequentially distribute control command data packets from the control command set to the corresponding field actuators. The distribution process utilizes the cloud platform's command gateway service to call the NB-IoT network operator's message push interface. The command data packets are routed via the operator's wireless access network and core network to the NB-IoT communication module integrated in the target field actuator. In some embodiments, the cloud platform's command gateway attaches a unique transaction identifier to each command data packet and initiates a timeout retransmission mechanism. After successfully receiving and parsing the command, the field actuator returns an acknowledgment frame carrying the same transaction identifier to the cloud platform via the NB-IoT network. If the cloud platform does not receive an acknowledgment frame within a set time, it retransmits the command data packet. In some embodiments, the execution of the command list supports condition-triggered logic, meaning that the execution of certain atomic control actions requires waiting for a completion confirmation signal from a preceding action or for specific sensor data to reach a preset threshold. It is understood that the field actuators are controllable devices such as electric valves, variable frequency pumps, and dosing devices installed at key nodes in the pipeline network. It is understandable that after the instruction is issued, the system enters a status monitoring cycle, collects new multi-dimensional real-time sensing data through the corresponding NB-IoT terminal device, and feeds the new multi-dimensional real-time sensing data back to the cloud platform to start a new round of management cycle.

[0040] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A smart water management cloud method driven by NB-IoT communication technology, characterized in that, The method includes: Collect multidimensional real-time sensor data on the operation status of the pipeline network, and obtain a standardized water status dataset based on the multidimensional real-time sensor data; Based on the preset pipeline network topology model and real-time operating parameters, a spatiotemporal correlation analysis is performed on the standardized water status dataset to generate a water status correlation diagram that reflects the dynamic correlation within the pipeline network. The water status association map is input into a pre-trained deep learning inference network. The deep learning inference network performs feature extraction and pattern recognition on the water status association map and outputs the inference results of potential abnormal events in the pipeline network and their probability distribution. Based on the inference results, the abnormal subgraph structure associated with potential abnormal events is located in the water status association diagram, and the root cause analysis is performed on the abnormal subgraph structure to determine the key monitoring node or pipeline section that caused the abnormality. Based on historical data from the key monitoring nodes or pipeline sections, several candidate control strategies are generated for the potential abnormal events. The execution effect of the candidate control strategies is simulated using a simulation environment. Based on the simulation results, the effectiveness of each candidate control strategy is evaluated and the risk is quantified, generating a strategy evaluation report.

2. The smart water management cloud method based on NB-IoT communication technology according to claim 1, characterized in that, The obtained standardized water status dataset includes: The multidimensional real-time sensing data is uploaded to the cloud platform via the NB-IoT network; The cloud platform receives the multidimensional real-time sensing data and performs format parsing and outlier cleaning on the data to obtain a standardized water status dataset, specifically including: The heterogeneous data formats generated by different models of NB-IoT terminal devices in the multidimensional real-time sensing data are identified, and the heterogeneous data formats are converted into a unified standardized data format within the cloud platform according to a preset format mapping table. After the sensor data stream is converted to a standardized data format, a time window-based integrity check is performed, and missing data points are marked and supplemented. Perform outlier detection based on statistical distribution and physical constraints on the sensor data stream with supplemented data points to identify outlier data points that deviate from the normal range or change pattern; For the identified abnormal data points, an interpolation algorithm based on the characteristics of its neighboring data points and historical data from the same period is used to repair the data or remove the flags. The sensor data stream, after data repair or label removal, is subjected to dimensional normalization to convert data values ​​of different physical dimensions into a unified numerical range, and finally outputs the standardized water status dataset.

3. The smart water management cloud method based on NB-IoT communication technology according to claim 2, characterized in that, Based on a preset pipeline network topology model and real-time operating parameters, the standardized water status dataset is subjected to spatiotemporal correlation analysis to generate a water status correlation diagram reflecting the dynamic correlation within the pipeline network, including: Load the pipeline topology model that describes the location of monitoring nodes, pipeline connections, and hydraulic properties in the pipeline network; Extract the status data of each monitoring node at the same time from the standardized water status dataset, and assign the status data as an attribute to the corresponding node in the pipeline topology model; Based on the real-time operating parameters, the weights of the edges connecting each node in the pipeline topology model are dynamically calculated, and the weights reflect the real-time strength of the hydraulic connection between the nodes. Construct a graph data structure with monitoring nodes as vertices and weighted edges representing the dynamic relationships between nodes as the initial relationship graph; Within a preset time window, the propagation path and delay effect of node state changes in the initial association graph are analyzed, the weights of the corresponding edges are strengthened or weakened, and finally the water affairs state association graph that can reflect the state propagation and dynamic coupling relationship is generated.

4. The smart water management cloud method based on NB-IoT communication technology according to claim 3, characterized in that, The water status correlation map is input into a pre-trained deep learning inference network. The deep learning inference network performs feature extraction and pattern recognition on the water status correlation map, and outputs the inference results of potential abnormal events in the pipeline network and their probability distribution, including: The attribute features of the nodes and the weight features of the edges in the water status association graph are jointly encoded into a high-dimensional feature tensor. The feature tensor is input into the graph convolutional layer of the deep learning inference network. The graph convolutional layer learns and extracts the local topological features of the water status association graph by aggregating the features of the node itself and its neighboring nodes. The extracted local topological features are input into the attention mechanism layer of the deep learning inference network. The attention mechanism layer calculates the importance weights of different nodes and edges in the graph to the overall state representation, and performs weighted fusion of features accordingly. The weighted and fused features are input into the classification inference layer of the deep learning inference network. The classification inference layer outputs the probability distribution of the water status association graph corresponding to various predefined abnormal event categories. The probability distribution is the inference result.

5. The smart water management cloud method based on NB-IoT communication technology according to claim 4, characterized in that, Based on the inference results, the abnormal subgraph structure associated with potential abnormal events is located in the water status correlation diagram, and root cause analysis is performed on the abnormal subgraph structure to determine the key monitoring nodes or pipeline sections that cause the anomalies, including: Extract potential abnormal event types with a probability exceeding a preset threshold from the inference results; In the water status association diagram, based on the feature patterns corresponding to the potential abnormal event types, a subgraph structure matching the feature patterns is searched, and the searched subgraph structure is marked as a candidate abnormal subgraph. For each candidate anomaly subgraph, the anomaly contribution of its internal node states is calculated, whereby the anomaly contribution quantifies the magnitude of the contribution of node state changes to the overall anomaly pattern. The nodes in the candidate anomaly subgraph are sorted according to the anomaly contribution degree, and the nodes with high contribution degrees are traced upstream or to the root source along the connection edges of the water status association graph. By combining the hydraulic transmission direction of the pipeline network with historical anomaly records, nodes that can serve as the origin of events or continuous pipe segments consisting of multiple nodes are identified in the candidate anomaly subgraph, and these are determined as the key monitoring nodes or pipeline segments.

6. The smart water management cloud method based on NB-IoT communication technology according to claim 5, characterized in that, Based on historical data from the key monitoring nodes or pipeline sections, several candidate control strategies are generated for the potential abnormal events, including: Retrieve the status data sequence of the key monitoring nodes or pipeline sections under similar past operating conditions and the records of the control actions taken from the historical database; Using the aforementioned pipeline physical model, the transmission effect on the hydraulic state of the entire pipeline network when different control actions are applied at the key monitoring nodes or pipeline sections is simulated. Based on the historical data sequence, the records of control actions, and the simulation results of the pipeline physical model, a strategy knowledge base is constructed. Based on the characteristics of the potential abnormal events, retrieve the basic strategy template from the strategy knowledge base; The retrieved basic strategy template is parametrically mutated and logically combined to generate multiple candidate control strategies that differ in control objectives, control intensity, and execution timing.

7. The smart water management cloud method based on NB-IoT communication technology according to claim 6, characterized in that, The process involves using a simulation environment to extrapolate the execution effects of the candidate control strategies, and then evaluating the effectiveness and quantifying the risks of each candidate strategy based on the extrapolation results, generating a strategy evaluation report, including: Construct a high-fidelity simulation environment based on the aforementioned pipeline network physical model and the current real-time water status data; Each of the candidate control strategies is converted into a sequence of instructions recognizable by the simulation environment and executed sequentially in the simulation environment; During execution, the trajectory of changes in key state variables in the simulation environment is recorded until the preset future time point is deduced. Based on the state of the pipeline network at the end of the simulation, multiple performance indicators for each candidate control strategy are calculated, including the degree of anomaly mitigation, energy consumption change, and recovery time. Identify secondary risk events that occur during the simulation process, and quantitatively assess the probability of occurrence and severity of impact for each identified secondary risk event; The quantitative evaluation results of all the effectiveness indicators and secondary risk events of each candidate regulation strategy are summarized and compiled into a structured strategy evaluation report.

8. The smart water management cloud method based on NB-IoT communication technology according to claim 7, characterized in that, The method includes: Based on the strategy evaluation report, an optimal control strategy is selected from the candidate control strategies using a multi-objective decision-making algorithm. The optimal control strategy is then decomposed into a specific set of control commands. The set of control commands is then sent to the field actuators corresponding to the key monitoring nodes or pipeline sections, driving the field actuators to execute the set of control commands to adjust the pipeline operation status. After the field actuator performs the control, it collects new multi-dimensional real-time sensing data through the corresponding NB-IoT terminal device and feeds the new multi-dimensional real-time sensing data back to the cloud platform to start a new round of management cycle; Based on the strategy evaluation report, an optimal control strategy is selected from the candidate control strategies using a multi-objective decision-making algorithm, including: Extract multiple performance indicators and risk quantification values ​​for each candidate control strategy from the strategy evaluation report; The performance index values ​​and risk quantification values ​​are normalized to eliminate the influence of different dimensions and uniformly converted into benefit-type or cost-type indicators. Based on the preference weights preset by the management side, a comprehensive evaluation function is constructed. The comprehensive evaluation function is the weighted sum of the normalized index values. Substitute the index value of each candidate regulation strategy into the comprehensive evaluation function to calculate its comprehensive score; Introduce constraints to exclude candidate control strategies that violate key physical limits or safety red lines; Among the candidate control strategies that meet the constraints, the strategy with the highest comprehensive score is selected as the optimal control strategy.

9. The smart water management cloud method based on NB-IoT communication technology according to claim 8, characterized in that, The optimal control strategy is decomposed into a specific set of control commands, and the set of control commands is issued to the field actuators corresponding to the key monitoring nodes or pipeline sections, including: The optimal control strategy is analyzed and broken down into a series of atomic control actions arranged in a time or logical order; For each atomic control action, match the field actuator of its target and determine the unique network identifier and control interface protocol of the field actuator; The parameters of each atomic control action are encapsulated into a specific instruction data packet according to the control interface protocol of the corresponding field actuator; All instruction data packets corresponding to atomic control actions are arranged into an ordered instruction list according to the execution sequence to form the control instruction set; Through the cloud platform and NB-IoT network, the instruction data packets in the control instruction set are sequentially sent to the corresponding field actuators.

10. A smart water management cloud system driven by NB-IoT communication technology, characterized in that: It includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the smart water management cloud method based on NB-IoT communication technology as described in any one of claims 1 to 9.

Images